The National AI Compute Scorecard

Four countries or blocs have national AI compute programmes large enough to compare. Each chose a different model. The metric each one publishes to describe its progress says more about the model than the programme's own claims do.

This piece is a scorecard of what can be verified from public sources, not a ranking of national intelligence. The United States routes donated compute to researchers. The United Kingdom bought two AIRR machines at Bristol and Cambridge, while a separate Edinburgh national system was cancelled and later restarted. The European Union selected nineteen AI Factory sites and thirteen antennas. India empaneled private cloud providers and subsidises access. China is left out of the four-way table because its public ledgers are not comparable in the same way. That is a limit of the method, not a claim that China is irrelevant.

The question under every section is simple. What did the programme promise, what did it spend or receive, and what did it publish as proof of use.

Why national compute became a race

From 2023 onward, governments treated access to advanced GPUs as a strategic input, close to energy or spectrum. Training frontier models needs dense clusters. Fine-tuning and inference need cheaper, more distributed capacity. Universities and startups cannot always buy that capacity at market rates. So states stepped in with pilots, factories, and missions.

The race language is easy. The accounting is hard. A GPU can be announced, ordered, installed, empaneled, assigned, or used. Those are six different states. Press releases often jump from the first to the last. A scorecard has to refuse that jump.

Three numbers matter for any national programme.

1. Cash or in-kind value actually transferred. Not the multi-year ceiling.
2. Capacity that exists. Machines installed, or capacity contractually available.
3. Use. Hours delivered to named users, or projects completed with those hours.

Most programmes publish a mix of (1) and (2). Almost none publish (3) in a form that lets a reader divide hours by public money.

The US. Donations routed to researchers

The United States did not buy a single GPU for its national programme. The National AI Research Resource pilot, launched in January 2024 under the National Science Foundation, coordinates donated compute from private companies and federal agencies and routes researchers and students to the machines.

The NSF NAIRR page, last updated in March 2026, estimates ~$100 million in private-sector in-kind contributions over the pilot's first two years. The largest contributions include NVIDIA ($30 million in compute and software), Microsoft ($20 million in Azure credits), OpenAI ($1 million in model access), and Cerebras (up to four EXAFLOPs of compute). 28 nongovernmental partners have contributed.

Fourteen federal agencies participate, from NASA and DARPA to the FDA and the Department of War. The Department of Energy and the National Institutes of Health co-lead a separate secure allocation track called NAIRR Secure for privacy-sensitive AI research. That split matters. Open research and sensitive research cannot share the same access rules. The US built the split into the programme rather than pretending one portal fits both.

The NAIRR pilot publishes a running tally on its portal, nairrpilot.org. As of August 2026 the portal listed 840 research projects, 85 NAIRR Classroom awards, 23 infrastructure and data demonstration projects, and 3 community workshops. The NSF page, last updated March 2026, cited more than 600 projects and 6,000 students across all 50 states, Washington, DC, and Puerto Rico. The gap between March and August figures is normal for a live portal. It also shows why a scorecard needs a verification date on every row.

The US model does not spend public money on GPUs for NAIRR. It does not build data centres under the NAIRR brand. It coordinates donations and routes researchers to them. Its primary output metric is projects supported. The metric reveals the model. The programme measures throughput, not owned capacity.

There are limits. In-kind dollars are estimated, not audited like appropriated spend. A $30 million NVIDIA contribution is not the same object as $30 million of cash on a Treasury line. Project counts can rise while hours per project stay thin. Classroom awards measure education reach, not model training. Still, among the four programmes in this piece, NAIRR is the only one whose public dashboard is organised around users and projects rather than machines and sites.

A second limit is durability. Donation programmes depend on corporate goodwill and surplus capacity. When private demand for GPUs is high, donated queues can shrink. The US bet is that coordination and secure tracks matter more than owning the silicon. That bet can be right for research access and wrong for industrial sovereignty at the same time.

The UK. Two AIRR machines built, a separate national system cancelled then restarted

The UK's AI Research Resource is a public procurement programme, not a donation portal. In the March 2023 Spring Budget, the government set aside £100 million for AIRR. On 1 November 2023, at the AI Safety Summit, that figure was tripled to £300 million for phase one, to build and connect two machines.

Those two machines exist. Isambard-AI at the University of Bristol uses 5,448 NVIDIA GH200 Grace Hopper superchips and about £225 million in public money. Dawn at the University of Cambridge is the second AIRR site. GOV.UK describes AIRR, launched in July 2025, as giving researchers and start-ups free access to both systems. A BBC report on Isambard-AI coming fully online also names Dawn as part of the same national resource.

A separate story often gets mixed into AIRR. The previous government planned a large national supercomputer in Edinburgh, talked about as an exascale system with funding in the hundreds of millions. The incoming Labour government cancelled or paused that plan after taking office, saying the money had never been formally allocated. In 2025 the government recommitted up to £750 million for a new national supercomputer at Edinburgh, now described as working alongside AIRR rather than as AIRR's missing second box. Treating Edinburgh as "AIRR machine number two" is the error this piece used to make. AIRR's phase-one pair was Bristol and Cambridge. Edinburgh is a different ledger line.

The UK also participates in the EU's AI Factory network as an antenna partner, listed as UKAIFA in the European Commission's AI Factories catalogue. Joining a European antenna after domestic budget churn is a policy choice, not proof that AIRR failed. Isambard-AI and Dawn are real systems.

The UK model is still straightforward public procurement. Buy machines, install them at universities, run them for researchers and, through later programmes, for start-ups. Its natural metrics would be machine utilisation, queue length, and hours by user class. It has published announcements, chip counts, and site names. It has not published utilisation in a form that sits cleanly next to the £300 million phase-one figure, the £225 million Isambard-AI figure, and the later Edinburgh pledge.

The EU. Nineteen sites, zero published utilisation hours

The EU's AI Factories are the best-documented of the four programmes on selection history. The European Commission publishes the full selection waves.

The network now stands at 19 AI Factories and 13 antennas across the EU and partner countries. The Commission says at least nine new AI-optimised supercomputers will be procured, which it says will more than triple the EuroHPC AI computing capacity. Total investment through the EuroHPC Joint Undertaking for the 2021-2027 period is €10 billion, counting EU, member state, and associated country money.

The EU has also opened a call for up to seven AI Gigafactories, facilities of more than 100,000 advanced AI processors each, supported by up to €10 billion in public funding and an estimated €20 billion in private investment. Gigafactories are a different scale from the factory network. Mixing the two in a single cheer line is a common source of confusion. Factories are selected sites and support programmes. Gigafactories are a later, larger procurement concept.

The EU publishes factory selections, site counts, and partner lists. The EuroHPC AI Factories page says the factories offer "free, customised support" to SMEs and startups. It does not publish running hours delivered to those startups in a comparable format. The factory count is a commitment metric. The hours would be a delivery metric. At 19 factories, the EU has published the most commitments. The delivery ledger is not on the same page.

Federalism explains part of this. EuroHPC money mixes EU and member-state contributions. Sites sit in different countries with different reporting habits. Harmonising utilisation across nineteen sites is harder than publishing a selection press release. Hard is not the same as impossible. Without hours, a reader cannot tell whether the network is busy, empty, or uneven.

The EU approach has a strength the others lack. Geographic spread. A researcher in a smaller member state can, in principle, reach a factory without emigrating to one capital. Whether that access is fast and real is exactly what utilisation and wait-time data would show.

India. Rented machines, counted but underused

India's programme is covered in full in the companion piece on this site. A condensed account belongs here so the four-way table is self-contained.

The model is empanelment rental. IndiaAI does not buy machines for a national warehouse. It empanels private cloud providers, approves users, and subsidises access. MeitY reported 38,000 GPUs onboarded by February 2026. Cash released by that date was ₹400.94 crore, 3.9% of the ₹10,371.92 crore outlay. Press reporting put utilised GPUs at 7,418 of 33,099 committed.

India's primary metric is GPUs onboarded. Its secondary metric, utilisation, is reported by a newspaper citing a letter to providers, not by a parliamentary utilisation annex. The PIB tender tables do publish L1 prices by model, which is more price transparency than most programmes offer. Price transparency without hours still leaves the scorecard incomplete.

The Indian design has a clear logic. Renting shifts idle-capacity risk toward providers. Subsidy can target approved users. The state avoids a large capital stock that may age badly. The same design makes "38,000 GPUs" a catalogue number. Readers who hear a national fleet are hearing a different object than the one in the contracts.

What each programme counts

None of the four publishes hours of GPU use per unit of public money in a clean, comparable table. The US publishes the metric closest to actual use. The EU publishes the most commitments. The UK built its phase-one pair and separately cancelled, then restarted, a larger Edinburgh national system. India publishes a count that a newspaper has to translate into use.

These mismatches track the models. The US cannot publish a GPU ownership count because it does not buy GPUs for NAIRR. It can publish projects supported. India can publish an onboarded count because it empanels providers and can count what they report. It has not published utilisation in the same official channel because utilisation comes from provider operations, not from a single owned cluster dashboard.

This is why cross-country "GPU races" in the press are often false contests. Comparing India's onboarded GPUs to the UK's installed Grace Hoppers to the EU's factory count to NAIRR's project awards is comparing four different units. A race needs a shared finish line. These programmes have not agreed on one.

The missing China row

Leaving China out needs an explicit reason. China has large public and quasi-public compute builds, provincial clusters, and industrial policy tools that dwarf several of the programmes above. The problem for this scorecard is verification symmetry. The four programmes here publish selection lists, parliamentary replies, or agency portals that an outside researcher can cite without relying on a single state media digest. Comparable primary series for Chinese national AI compute, with cash, capacity, and hours in one frame, were not available in the same form during the August 2026 verification pass.

That absence biases the table toward open reporting systems. It does not prove those systems deliver more compute. Add China when a primary ledger of similar quality appears. Until then, treat the four-way comparison as a study of transparent programmes, not a map of global capacity.

How the models allocate risk

Each model places downside in a different place.

• US donation. Risk sits with corporate donors and with researchers who depend on queues that can shrink. The public budget risk is low. The sovereignty risk is high if donations pause.
• UK purchase. Risk sits with the public balance sheet for capital and with universities for operations. Political risk shows up when parallel national systems, such as Edinburgh, are paused and restarted. Utilisation risk is operational.
• EU co-federation. Risk is shared across member states and the Joint Undertaking. Coordination risk is high. Geographic access can be a strength if hours are real.
• India rental. Risk of idle hardware sits more with providers. Fiscal risk sits with subsidy releases. Catalogue metrics can outrun use.

None of these risk maps is automatically better. A country short on capital may prefer rental. A country that wants owned capacity for defence research may prefer purchase. A country with deep private surplus may prefer donation. The scorecard's job is to match the claim to the model, then ask for the metric the model can honestly produce.

What a fair comparison would look like in practice

Imagine four finance ministries asked to fill the same spreadsheet once a quarter.

A country that cannot fill a row should leave it blank, not replace it with a press metric. Blank rows are information. They tell the reader where the ledger stops.

NAIRR could fill project counts and partner in-kind estimates today. India could fill tender prices and cash released today. The EU could fill site counts today. The UK could fill installed chips at Isambard-AI today. Hours and wait times are the shared blank across all four.

Education, industry, and defence are not one queue

National programmes often sell one portal and serve three missions.

Education wants classroom access, short jobs, and wide geographic reach. Industry wants predictable queues, commercial terms, and sometimes confidential workloads. Defence and sensitive research want air-gapped or tightly controlled tracks.

The US built NAIRR Secure as a separate track for privacy-sensitive work. That is an honest design choice. A single public dashboard that mixes classroom awards with secure allocations would mislead. India's user table from Parliament mixes researchers, startups, students, and government slots. That is useful and still coarse. The EU's SME support language sits next to national supercomputing cultures that grew up serving science first. The UK's university-hosted machine inherits an academic operating model by default.

When a minister says the programme serves startups, the scorecard should ask what share of hours went to startups last quarter. When a minister says it serves students, the scorecard should ask how many student hours cleared the queue. Mission mix without hour mix is a speech, not a ledger.

Energy and siting sit under every GPU count

A national compute claim that ignores power is incomplete. Clusters need megawatts, transformers, and often new substations. Announcing accelerators without interconnection dates repeats the financing error from the sovereign debt story on this site. Capacity on a slide is not capacity on a busbar.

The EU's Gigafactory concept, facilities above 100,000 advanced processors, makes the power problem impossible to hide. India's rental model partly outsources siting to cloud providers who already hold power contracts. The US donation model rides on private facilities that already exist. The UK's single-machine approach concentrates the siting problem at Bristol. Each model handles power differently. None of the four public scorecards currently lead with megawatts delivered to AI workloads.

A future row in the shared spreadsheet should be megawatt-hours consumed by the public programme's workloads. That number is hard. It is also closer to physical reality than onboarded GPU counts.

How press coverage distorts the race

Press incentives favour round numbers and national pride. ₹10,372 crore. £300 million. €10 billion. 38,000 GPUs. 19 factories. 840 projects. Each number is citeable. Few articles divide hours by money. Fewer still keep verification dates visible when a portal updates.

This distortion is not only the press's fault. Programmes choose metrics that travel well in headlines. Factory counts travel. Project counts travel. Onboarded GPUs travel. Wait-time histograms do not travel. Until funders reward the dull metrics, the dull metrics will stay unpublished.

One way to cut through the noise is a hard rule. No cross-country rank from mismatched units. No success claim from a ceiling. No utilisation claim from a catalogue. That rule alone removes most of the false race.

Objections that still stand

A defender of the current programmes can make several fair points.

First, pilots are allowed to be messy. NAIRR is a pilot. IndiaAI is early in a five-year window. EU factories are still being stood up. Demanding mature utilisation ledgers on day one can punish programmes that chose transparency over delay.

Second, some use is intentionally unpublished. Secure tracks, commercial confidentiality, and national-security workloads will never appear in a public hour table. A public scorecard will always undercount real activity.

Third, private cloud capacity may matter more than public programmes for most startups. A company with a cloud credit from a hyperscaler may never join NAIRR or IndiaAI. Measuring only public programmes can miss where work actually runs.

All three points are real. They do not erase the value of a public ledger for public money. They set the ceiling on what a public ledger can claim.

Shared metrics that would make the numbers comparable

GPU-hours delivered, divided by cash released or audited in-kind value, using the same formula in every country. Published tender or partner prices after subsidy, by model class. Wait time from application to first access, published as a distribution rather than an anecdote.

The US programme is closest on the first family of numbers because it counts projects and its partners' contributions are public. The UK is hardest to pin to one closed money ledger, because phase-one AIRR cash, the £225 million Isambard-AI figure, later AIRR expansion pledges, and the Edinburgh national system sit in overlapping announcements. The EU is strongest on site documentation and weakest on hours. India is strongest on tender price tables and weakest on official utilisation.

Until those shared metrics appear, the national AI compute race is scored on announcements, and the announcement that moves fastest is the next one.

Sources and limits

NAIRR figures come from nairrpilot.org and nsf.gov. EU figures come from digital-strategy.ec.europa.eu and eurohpc-ju.europa.eu. UK figures come from GOV.UK and UKRI announcements for AIRR phase one, University of Bristol materials for Isambard-AI, and later GOV.UK notes on Dawn and the Edinburgh national system. India figures come from the IndiaAI piece on this site, which traces the Rajya Sabha annex and PIB releases. Verification date August 2026.

This piece does not estimate private cloud capacity inside each country. Private capacity often dwarfs public programmes. Amazon and Microsoft alone have committed tens of billions to Indian cloud and AI infrastructure, as noted in the IndiaAI companion. A national scorecard of public programmes is still useful. It answers what taxpayers and public agencies can verify. It does not answer who owns the largest share of silicon on the ground.